A watermark is evidence, not a verdict

A text watermark can answer a useful question: is this passage statistically consistent with text produced or substantially edited by a particular AI model?

It cannot tell us who authored the final work, whether someone intended to deceive, or what consequence should follow. That distinction will matter as watermark detection moves from a technical capability into schools, workplaces, publishing, investigations, and public policy.

On August 14, Anthropic announced that future Claude models will generate watermarked text. The company plans to apply the watermark globally and add it to older models over the coming months. A detection API is promised, but its design and availability have not yet been published.

The technology is arriving. The rules for using its results should arrive first.

What the watermark actually says

Anthropic's approach is based on Google DeepMind's SynthID-Text method. Language models already choose among several plausible next words. Watermarking subtly changes the source of randomness behind those choices, leaving a statistical pattern across a passage. A detector with the corresponding key can then estimate the likelihood that Claude was involved in producing the text.

Nothing visible is added. There are no hidden characters, extra tokens, or user identifiers. Anthropic says the watermark cannot be traced to a person, organization, or conversation.

That privacy boundary is welcome, but the evidentiary boundary is just as important. A positive result does not establish whether Claude wrote a passage from scratch or heavily edited it. It does not identify the human who submitted or revised the text. It says nothing about ownership, responsibility, or whether the work was permitted.

Anthropic connects the change to the EU AI Act and its voluntary Code of Practice on Transparency of AI-generated Content. The European Commission says the related Article 50 transparency obligations became applicable on August 2, 2026. The code offers signatories a practical path for marking and detecting generated content, but it does not make every detection result conclusive.

Confidence depends on context

Watermark detection becomes more reliable when there is more varied model-written text to examine. Its limits appear where language offers fewer genuine choices.

Short samples contain less signal. Factual passages and code often have a narrow set of correct next tokens. Proofreading may change too few words to leave a detectable pattern. Light editing may preserve a watermark, while a complete rewrite can remove it. A provider's key can test for that provider's watermark; it cannot prove that unmarked text was written by a person rather than another model.

These are not minor footnotes. They determine what a result can reasonably support. A confidence estimate detached from sample length, content type, editing history, model version, and detection method is easy to overread.

Decide what the signal may do

Before adopting a detector, an organization should define its permitted uses. Using a watermark result to prioritize a review is very different from using it to reject an applicant, discipline an employee, accuse a student, remove a publication, or make a fraud determination.

For any consequential use, the operating policy should answer five questions:

1. What exact decision may the watermark inform, and which decisions are off-limits?
2. What minimum sample and confidence conditions make the result usable?
3. What independent evidence must corroborate it?
4. Who can challenge the result, and who reviews that challenge?
5. What text, detector output, and contextual data will be retained, for how long, and with whose access?

The record should preserve the provider and model version, detector version, sample length, content type, known transformations, result, confidence, supporting evidence, final decision, and reviewer. Without that context, a later investigator may see a number but have no reliable way to understand it.

The PrivacyFirst view

The greatest risk is not that text watermarking is useless. It is that a useful, narrow signal will be promoted into a verdict because it arrives with a probability attached.

Good governance keeps the signal smaller than the decision. Watermark detection can strengthen provenance evidence when its limits, permitted uses, corroboration, retention, and appeal path are explicit. Used alone, it can turn uncertainty into false confidence.

The best time to write those rules is before the detector is switched on.